claude-graph provides 6 well-named, verb-prefixed tools with clear descriptions (avg 156 chars, within baseline 194). All tools have input schemas with typed parameters and descriptions. However, output schemas are not documented, responses are inferred from code (dict/list returns without explicit schema declarations). Error handling is minimal; no recovery guidance or actionable error messages. Parameter descriptions lack format constraints (e.g., 'pattern' enum values are documented in description text, not as JSON Schema enum). Missing tool annotations (readOnlyHint, destructiveHint). Composition is sound, tools chain well (e.g., search_nodes_tool → query_graph_tool). No security issues detected (no secrets in params, no destructive operations without safeguards).

• Add explicit JSON Schema output schemas to all tools. Document return types with typed fields (e.g., 'Returns: {files: int, nodes: int, edges: int, languages: [str]}'). This enables LLMs to plan multi-step workflows.
• Convert 'pattern' and 'scope' parameter descriptions to JSON Schema enum constraints. Replace 'One of: callers_of, callees_of, ...' with enum: ['callers_of', 'callees_of', ...] in the schema.
• Add error handling with actionable recovery guidance. E.g., if database doesn't exist, return {error: 'Graph not built. Call build_or_update_graph() first.', recovery: 'build_or_update_graph'}.
• Remove '_tool' suffix from tool names. Rename query_graph_tool → query_graph, get_impact_radius_tool → get_impact_radius, search_nodes_tool → search_nodes, render_graph_tool → render_graph.
• Add tool annotations: mark build_or_update_graph and render_graph_tool with destructiveHint=true; mark all read-only tools with readOnlyHint=true. This guides agent retry logic.